Vision QC on the line: the last five percent is economics
A packaging line runs at three hundred units a minute. A machining cell turns out a part every eleven seconds. At that pace a human inspector is not inspecting. They are sampling, and hoping the sample is honest. Vision quality control exists because the arithmetic of sampling stops working the moment the cost of an escaped defect gets serious.
We build vision systems for manufacturers, and the pattern of what works has become clear enough to write down. The figures in this piece are illustrative of what we see across plants, not audited results from a named client. They are here to make the economics concrete, because vision QC is decided on economics, not on model benchmarks.
What the camera is actually for
The camera does not replace the inspector. It replaces the part of the inspector's job that humans are worst at: doing the same visual check, to the same standard, thousands of times, without drifting. A person is good at judgement and terrible at repetition. A model is the reverse. The systems that survive divide the work along exactly that line.
In practice that means the camera checks every unit, not a sample. Presence of a weld. Position of a label. Fill level. Seal integrity. Surface scratches above a defined size. The confident passes go straight through. The confident failures get rejected with the image attached, so nobody argues later about what the camera saw. The narrow band in the middle, the borderline calls, goes to a person. That band is where the design lives or dies, and we will come back to it.
The cost of a defect depends on where you catch it
One curve drives the whole business case: the later a defect is caught, the more it costs. Catch it at the station and you scrap one part. Catch it at end of line and you have added value to a part you are about to bin, and you get to go hunting for its siblings. Catch it in finished goods and there is a quarantine, a re-inspection of the batch, and a delayed shipment. Catch it at the customer and you are paying for freight, returns, credit notes, and a buyer who now audits you harder. The multiples below are illustrative, but the shape holds in nearly every plant we have seen.
This curve is why a merely decent vision system placed early often beats an excellent one placed late. The value is not in the accuracy number on its own. It is in position and coverage: every unit checked, at the point where a reject costs the least, by something that does not get tired at hour seven of a shift.
Consistency is the quiet win
Ask a quality manager what actually worries them and it is rarely the average defect rate. It is the variance. The day shift and the night shift disagree about what counts as a scratch. The working definition of acceptable drifts between inspectors, and within a single inspector as the shift wears on. Every re-inspection exercise we have seen tells the same story: people disagree with each other, and with their own earlier calls.
The camera's number is not the headline. The headline is that it is the same number at 3am on a Sunday, on the last unit before changeover, at the same threshold it applied in January. Consistency is what lets you tighten the standard and mean it, and it is what makes the quality data trustworthy enough to act on.
Borderline calls belong to people
No production model is confidently right about everything, and pretending otherwise is how vision projects die. The honest design routes by confidence. Above a threshold, the system acts on its own. Below it, the unit is held and a person decides, with the image, the measurement, and the spec side by side on one screen. Those human calls feed back into the training set, so the borderline band narrows over months instead of being argued about in the corridor.
This also changes the inspector's day. Instead of staring at a thousand good parts to find three bad ones, they look at forty genuinely ambiguous cases and apply real judgement to each. That is a better job, and in our experience it is why the people who were expected to resist the system often end up defending it.
Traceability is half the value
Every automated inspection produces a record: the image, the decision, the threshold in force, the timestamp, the batch and lot. Kept properly, that record answers the questions that used to take days of digging. Which lots did the bad roll of film touch. When did this scratch pattern start, and which tool change does it line up with. What exactly did we ship to the customer who is complaining, and can we prove it left here clean. A vision system without traceability is a very expensive reject gate. With it, the line gets a memory.
The last five percent is an economics question
Every vision project reaches the same meeting. The system sits at, say, 94 percent agreement with the standard, and someone asks what it takes to reach 99. The honest answer is: considerably more than the first 94 cost. Rare defects need rare examples, so you wait for them or stage them. Lighting gets re-engineered. Sometimes the fix is new optics, sometimes it is a physical change to how parts present to the camera. The curve is steep, and almost none of it is model work. It is data collection and line engineering, and it has a price.
Whether that price is worth paying is arithmetic, not ambition. If one escaped defect costs you real money at the customer, the last points of accuracy can pay for themselves in a single avoided incident. If escapes are cheap and rework is easy, then 94 percent with a person on the borderline calls is the correct place to stop. The model did not make that decision. The defect cost curve did, and it should be on the table when you make it.
The last five percent of accuracy is not a modelling question. It is a purchasing decision, and it should be made with the defect cost curve on the table.